Domain Alignment Dynamic Spectral and Spatial Feature Fusion for Hyperspectral Change Detection

Change detection is an important task in geospatial analysis that aims to identify noticeable variations in geographic elements between images captured at different periods. However, existing methods often overlook the distribution discrepancies across images caused by changes in imaging time. Meanwhile, the spectral and spatial features of hyperspectral images still have great potential for further development in extracting and detecting changes. To mitigate these challenges, we propose a novel approach called domain alignment dynamic spectral and spatial feature fusion (DADSSFF) for hyperspectral change detection. First, DADSSFF uses the main network to optimize the alignment of the mean (first-order statistics) and correlation (variance, second-order statistics) of the bitemporal images, coordinating features across both levels to alleviate the issue of inconsistent feature distribution. Second, the Kullback–Leibler divergence is employed to increase the interaction between the two auxiliary networks and the main network, enhancing the extraction of spectral and spatial attention features from bitemporal hyperspectral images. Finally, the cosine similarity is applied to measure the weights of the spectral and spatial features, enabling a dynamic evaluation of their importance. The effectiveness of DADSSFF is demonstrated by experimental results on three classical hyperspectral change detection datasets.